Hainich with ERA-Interim

Manual fine tuning learning process for presentation 18 Jan 2023

%load_ext autoreload
%autoreload 2

from meteo_imp.kalman.fastai import *
from meteo_imp.kalman.filter import *
from meteo_imp.utils import *
from meteo_imp.data import  *
from meteo_imp.data import  _def_meteo_vars, units

from fastai.tabular.learner import *
from fastai.learner import *
from fastai.callback.all import *
from fastcore.foundation import L

from meteo_imp.kalman.fastai import show_results
import pandas as pd
import numpy as np
import torch
from pyprojroot import here

from sklearn.decomposition import PCA

base_dir = here("analysis/presentations/plots_18_jan")
base_dir.mkdir(exist_ok=True)

def save_plot(p, path):
    f_name = base_dir / (path + ".vl.json")
    with open(f_name, 'w') as f:
        f.write(p.to_json())
    return f_name

reset_seed()

hai = pd.read_parquet(hai_path64)
hai_era = pd.read_parquet(hai_era_path64)

PCA

pc = PCA().fit(hai)

d0 = hai.iloc[0:1]

tr0 = pc.transform(d0)

tr0

array([[-121.11917652,   -7.06844313,    0.87975241]])

t0 = np.random.randn(5,1) * 10
tt = np.hstack([t0, 2 * t0 + np.random.randn(5,1), 3 * t0 + np.random.randn(5,1)])

tt = np.vstack([tt, -tt])

tt.mean(0)

array([1.77635684e-16, 0.00000000e+00, 0.00000000e+00])

def pca(X):
    # Data matrix X, assumes 0-centered
    n, m = X.shape
    assert np.allclose(X.mean(axis=0), np.zeros(m))
    # Compute covariance matrix
    C = np.dot(X.T, X) / (n-1)
    # Eigen decomposition
    eigen_vals, eigen_vecs = np.linalg.eig(C)
    # Project X onto PC space
    X_pca = np.dot(X, eigen_vecs)
    return X_pca, eigen_vecs

pca(tt)

(array([[-9.15947487e+00, -3.03773243e-01,  3.81626348e-01],
        [ 2.00045215e+01, -1.99595583e-02,  1.53745860e+00],
        [ 1.33744310e+01,  2.01395206e-02, -3.19819214e-02],
        [-8.99057811e+00,  1.24208069e-01,  9.57988123e-01],
        [ 3.14616875e+01, -4.88141095e-02, -5.79117681e-01],
        [ 9.15947487e+00,  3.03773243e-01, -3.81626348e-01],
        [-2.00045215e+01,  1.99595583e-02, -1.53745860e+00],
        [-1.33744310e+01, -2.01395206e-02,  3.19819214e-02],
        [ 8.99057811e+00, -1.24208069e-01, -9.57988123e-01],
        [-3.14616875e+01,  4.88141095e-02,  5.79117681e-01]]),
 array([[-0.26545847, -0.95659856,  0.12021234],
        [-0.53518386,  0.04249433, -0.84366608],
        [-0.80194142,  0.28829401,  0.52323659]]))

sk_pc = PCA(2).fit(tt)

tt @ sk_pc.components_.T

array([[ -9.15947487,   0.38162635],
       [ 20.00452148,   1.5374586 ],
       [ 13.37443103,  -0.03198192],
       [ -8.99057811,   0.95798812],
       [ 31.46168754,  -0.57911768],
       [  9.15947487,  -0.38162635],
       [-20.00452148,  -1.5374586 ],
       [-13.37443103,   0.03198192],
       [  8.99057811,  -0.95798812],
       [-31.46168754,   0.57911768]])

tr = sk_pc.transform(tt)
tr

array([[ -9.15947487,   0.38162635],
       [ 20.00452148,   1.5374586 ],
       [ 13.37443103,  -0.03198192],
       [ -8.99057811,   0.95798812],
       [ 31.46168754,  -0.57911768],
       [  9.15947487,  -0.38162635],
       [-20.00452148,  -1.5374586 ],
       [-13.37443103,   0.03198192],
       [  8.99057811,  -0.95798812],
       [-31.46168754,   0.57911768]])

sk_pc.components_.T

array([[-0.26545847,  0.12021234],
       [-0.53518386, -0.84366608],
       [-0.80194142,  0.52323659]])

(sk_pc.components_ @ tt.T).T

array([[ -9.15947487,   0.38162635],
       [ 20.00452148,   1.5374586 ],
       [ 13.37443103,  -0.03198192],
       [ -8.99057811,   0.95798812],
       [ 31.46168754,  -0.57911768],
       [  9.15947487,  -0.38162635],
       [-20.00452148,  -1.5374586 ],
       [-13.37443103,   0.03198192],
       [  8.99057811,  -0.95798812],
       [-31.46168754,   0.57911768]])

tt[0, None]

array([[2.76792539, 4.56712931, 7.45746711]])

sk_pc.components_.shape

(3, 3)

tt[0].shape

(3,)

(sk_pc.components_ @ tt[0])

array([-9.15947487,  0.38162635])

tt[0]

array([2.76792539, 4.56712931, 7.45746711])

sk_pc.components_.T @ tr[0]

array([2.47733634, 4.58003795, 7.54504311])

sk_pc.inverse_transform(tr[0])

array([2.47733634, 4.58003795, 7.54504311])

d0.to_numpy()

array([[-0.6  ,  0.   ,  0.222]])

hai.iloc[0]

TA      -0.600
SW_IN    0.000
VPD      0.222
Name: 2000-01-01 00:30:00, dtype: float64

\[ x = y\Lambda \]

pc.components_

array([[ 0.01681572,  0.99979324,  0.01143269],
       [ 0.93010891, -0.01983747,  0.36674772],
       [-0.36689868, -0.00446652,  0.93025018]])

np.linalg.inv(pc.components_)

array([[ 0.01681572,  0.93010891, -0.36689868],
       [ 0.99979324, -0.01983747, -0.00446652],
       [ 0.01143269,  0.36674772,  0.93025018]])

Setup

assuming that the control has the same dimensions of the observations then if we are doing a local slope model we have \(B \in \mathbb{R}^{state \times contr}\): \[ B = \begin{bmatrix} -I & I \\ 0 & 0 \end{bmatrix}\]

from torch import hstack, eye, vstack, ones, zeros, tensor
from functools import partial

def set_dtype(*args, dtype=torch.float64):
    return [partial(arg, dtype=dtype) for arg in args] 

eye, ones, zeros, tensor = set_dtype(eye, ones, zeros, tensor)

def init_smart(n_dim_obs, n_dim_state, df, pca=True):
    # n_dim_obs == n_dim_contr
    if pca:
        comp = PCA(n_dim_state).fit(df).components_
        obs_matrix = tensor(comp.T) # transform state -> obs
        contr_matrix = tensor(comp) # transform obs -> state
    else:
        obs_matrix, contr_matrix = eye(n_dim_obs), eye(n_dim_obs)
        
    return KalmanFilter(
        trans_matrix =     vstack([hstack([eye(n_dim_state),                eye(n_dim_state)]),
                                   hstack([zeros(n_dim_state, n_dim_state), eye(n_dim_state)])]),
        trans_off =        zeros(n_dim_state * 2),        
        trans_cov =        eye(n_dim_state * 2)*.1,        
        obs_matrix =       hstack([obs_matrix, zeros(n_dim_obs, n_dim_state)]),
        obs_off =          zeros(n_dim_obs),          
        obs_cov =          eye(n_dim_obs)*.01,            
        contr_matrix =     vstack([hstack([-contr_matrix,                  contr_matrix]),
                                   hstack([ zeros(n_dim_state,n_dim_obs), zeros(n_dim_state, n_dim_obs)])]),
        init_state_mean =  zeros(n_dim_state * 2),        
        init_state_cov =   eye(n_dim_state * 2) * 3,
    ) 

np.hstack([np.eye(2), np.eye(2)])

array([[1., 0., 1., 0.],
       [0., 1., 0., 1.]])

init_smart(3,2, hai)

state	x_0	x_1	x_2	x_3
x_0	1.0000	0.0000	1.0000	0.0000
x_1	0.0000	1.0000	0.0000	1.0000
x_2	0.0000	0.0000	1.0000	0.0000
x_3	0.0000	0.0000	0.0000	1.0000

trans cov (Q)

state	x_0	x_1	x_2	x_3
x_0	0.1000	0.0000	0.0000	0.0000
x_1	0.0000	0.1000	0.0000	0.0000
x_2	0.0000	0.0000	0.1000	0.0000
x_3	0.0000	0.0000	0.0000	0.1000

trans off

state	offset
x_0	0.0000
x_1	0.0000
x_2	0.0000
x_3	0.0000

obs matrix (H)

variable	x_0	x_1	x_2	x_3
y_0	0.0168	0.9301	0.0000	0.0000
y_1	0.9998	-0.0198	0.0000	0.0000
y_2	0.0114	0.3667	0.0000	0.0000

obs cov (R)

variable	y_0	y_1	y_2
y_0	0.0100	0.0000	0.0000
y_1	0.0000	0.0100	0.0000
y_2	0.0000	0.0000	0.0100

obs off

variable	offset
y_0	0.0000
y_1	0.0000
y_2	0.0000

contr matrix (B)

state	c_0	c_1	c_2	c_3	c_4	c_5
x_0	-0.0168	-0.9998	-0.0114	0.0168	0.9998	0.0114
x_1	-0.9301	0.0198	-0.3667	0.9301	-0.0198	0.3667
x_2	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
x_3	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000

init state mean

state	mean
x_0	0.0000
x_1	0.0000
x_2	0.0000
x_3	0.0000

init state cov

state	x_0	x_1	x_2	x_3
x_0	3.0000	0.0000	0.0000	0.0000
x_1	0.0000	3.0000	0.0000	0.0000
x_2	0.0000	0.0000	3.0000	0.0000
x_3	0.0000	0.0000	0.0000	3.0000

class PersistentRecorder(Callback):
    order = 70
    name = "per_recorder"
    attrs = ['lrs', 'iters', 'losses', 'values']
    def before_fit(self):
        "Prepare state for training"
        for attr in self.attrs:
            if not hasattr(self.per_recorder, attr): setattr(self.per_recorder, attr, [])

    def after_batch(self):
        for attr in self.attrs:
            setattr(self.per_recorder, attr, getattr(self.recorder, attr))

models = []

dls = imp_dataloader(hai, hai_era, var_sel = ['TA', 'SW_IN', 'VPD'], block_len=200, gap_len=10, bs=20, control_lags=[1], n_rep=2)

len(dls.valid.items)
items = [dls.valid.items[i] for i in [10, 50, 100, 200, 300, 400]]

def train_show_save(learn, n_iter, lr):
    learn.fit(n_iter, lr)
    models.append(learn.model.state_dict().copy())
    learn.recorder.plot_loss()
    items = [learn.dls.valid.items[i] for i in [10, 50, 100, 200, 300, 400]]
    return show_results(learn, items = items, control=hai_control)

Gaps all variables

model = init_smart(3,3, hai, pca=False).cuda()
model.var_names = _def_meteo_vars.values()

loss = loss_func=KalmanLoss(only_gap=False)
learn = Learner(dls, model, loss, cbs=[Float64Callback], metrics=imp_metrics)

learn.model.use_smooth = True

show_results(learn, items=items, control=hai_control)

train_show_save(learn, 1, 1e-3)

epoch	train_loss	valid_loss	rmse	rmse_gap	r2	r2_gap	time
0	-87.370314	-96.212915	0.069196	0.210351	0.946861	-250571814513532732431918956544.000000	02:37

state	x_0	x_1	x_2	x_3	x_4	x_5
x_0	1.1489	0.0042	0.0120	0.8404	-0.0794	-0.1127
x_1	-0.0954	1.1464	-0.0546	0.1177	0.8560	0.0599
x_2	-0.0225	0.0420	1.1357	-0.0717	-0.1190	0.8475
x_3	0.1201	0.0661	0.1170	1.1840	-0.1150	-0.1073
x_4	0.0770	0.1299	0.0068	-0.1847	1.0987	-0.1500
x_5	0.1084	0.0358	0.1133	-0.1809	-0.0281	1.1533

trans cov (Q)

state	x_0	x_1	x_2	x_3	x_4	x_5
x_0	0.0341	0.0493	0.0361	-0.0008	-0.0312	0.0026
x_1	0.0493	0.1125	0.0531	0.0475	0.0047	0.0509
x_2	0.0361	0.0531	0.0399	0.0001	-0.0327	0.0058
x_3	-0.0008	0.0475	0.0001	0.0608	0.0614	0.0585
x_4	-0.0312	0.0047	-0.0327	0.0614	0.0901	0.0554
x_5	0.0026	0.0509	0.0058	0.0585	0.0554	0.0594

trans off

state	offset
x_0	0.0008
x_1	-0.0083
x_2	0.0067
x_3	0.0023
x_4	-0.0007
x_5	0.0005

obs matrix (H)

variable	x_0	x_1	x_2	x_3	x_4	x_5
y_0	0.8238	-0.1897	-0.1014	-0.0841	0.1914	0.1286
y_1	0.0814	0.8353	0.1048	0.1040	0.0127	0.0689
y_2	-0.0742	-0.1502	0.8296	0.1282	0.1811	-0.0746

obs cov (R)

variable	y_0	y_1	y_2
y_0	0.0084	0.0000	0.0000
y_1	0.0000	0.0084	0.0000
y_2	0.0000	0.0000	0.0084

obs off

variable	offset
y_0	0.0017
y_1	-0.0046
y_2	0.0026

contr matrix (B)

state	c_0	c_1	c_2	c_3	c_4	c_5
x_0	-1.0290	-0.0485	-0.0022	0.9297	0.0792	-0.0247
x_1	-0.0267	-0.8778	-0.0233	0.0113	0.9099	0.0261
x_2	-0.0018	-0.0592	-1.0044	-0.0310	0.0674	0.9204
x_3	-0.0503	-0.0312	-0.0556	-0.0440	-0.0529	-0.0587
x_4	-0.0375	-0.0537	-0.0246	-0.0412	-0.0352	-0.0338
x_5	-0.0653	-0.0179	-0.0575	-0.0655	-0.0307	-0.0449

init state mean

state	mean
x_0	-0.0393
x_1	0.0044
x_2	-0.0092
x_3	0.0207
x_4	-0.0094
x_5	-0.0021

init state cov

state	x_0	x_1	x_2	x_3	x_4	x_5
x_0	2.5500	-0.1402	-0.0082	0.5233	0.3011	0.0882
x_1	-0.1402	2.4578	0.1502	0.0355	0.5714	-0.0459
x_2	-0.0082	0.1502	2.5045	0.2739	0.0634	0.5195
x_3	0.5233	0.0355	0.2739	2.3814	-0.2201	-0.1842
x_4	0.3011	0.5714	0.0634	-0.2201	2.3880	-0.1162
x_5	0.0882	-0.0459	0.5195	-0.1842	-0.1162	2.3874

train_show_save(learn, 1, 1e-3)

/home/simone/anaconda3/envs/data-science/lib/python3.10/site-packages/fastai/callback/core.py:69: UserWarning: You are shadowing an attribute (__class__) that exists in the learner. Use `self.learn.__class__` to avoid this
  warn(f"You are shadowing an attribute ({name}) that exists in the learner. Use `self.learn.{name}` to avoid this")

epoch	train_loss	valid_loss	rmse	rmse_gap	r2	r2_gap	time
0	-108.023250	-113.113034	0.073622	0.198386	0.956828	-105578612000578019271909572608.000000	02:37

state	x_0	x_1	x_2	x_3	x_4	x_5
x_0	1.1440	-0.0057	0.0107	0.8262	-0.0845	-0.1205
x_1	-0.0534	1.2089	-0.0223	0.0948	0.7937	0.0494
x_2	-0.0289	0.0176	1.1417	-0.0702	-0.1154	0.8351
x_3	0.1277	0.0589	0.1074	1.1787	-0.1203	-0.1024
x_4	0.0654	0.1560	0.0043	-0.1781	1.1087	-0.1415
x_5	0.0959	0.0439	0.1197	-0.1725	-0.0340	1.1496

trans cov (Q)

state	x_0	x_1	x_2	x_3	x_4	x_5
x_0	0.0207	0.0254	0.0210	-0.0052	-0.0232	-0.0033
x_1	0.0254	0.0548	0.0255	0.0259	0.0034	0.0281
x_2	0.0210	0.0255	0.0222	-0.0056	-0.0241	-0.0023
x_3	-0.0052	0.0259	-0.0056	0.0473	0.0504	0.0465
x_4	-0.0232	0.0034	-0.0241	0.0504	0.0695	0.0473
x_5	-0.0033	0.0281	-0.0023	0.0465	0.0473	0.0477

trans off

state	offset
x_0	0.0004
x_1	-0.0055
x_2	0.0048
x_3	0.0027
x_4	-0.0033
x_5	0.0021

obs matrix (H)

variable	x_0	x_1	x_2	x_3	x_4	x_5
y_0	0.7902	-0.2032	-0.1217	-0.0986	0.2164	0.1090
y_1	0.0160	0.7417	0.0555	0.1016	0.1041	0.0374
y_2	-0.1103	-0.1594	0.7916	0.1185	0.2096	-0.0949

obs cov (R)

variable	y_0	y_1	y_2
y_0	0.0077	0.0000	0.0000
y_1	0.0000	0.0077	0.0000
y_2	0.0000	0.0000	0.0077

obs off

variable	offset
y_0	-0.0009
y_1	-0.0053
y_2	-0.0021

contr matrix (B)

state	c_0	c_1	c_2	c_3	c_4	c_5
x_0	-1.0277	-0.0595	-0.0019	0.9181	0.0838	-0.0194
x_1	-0.0414	-0.8661	-0.0323	0.0110	0.8741	0.0262
x_2	-0.0029	-0.0647	-1.0045	-0.0319	0.0757	0.8957
x_3	-0.0678	-0.0368	-0.0564	-0.0559	-0.0610	-0.0639
x_4	-0.0374	-0.0690	-0.0209	-0.0485	-0.0406	-0.0354
x_5	-0.0591	-0.0267	-0.0753	-0.0633	-0.0407	-0.0555

init state mean

state	mean
x_0	-0.0396
x_1	-0.0009
x_2	-0.0049
x_3	0.0153
x_4	-0.0139
x_5	-0.0094

init state cov

state	x_0	x_1	x_2	x_3	x_4	x_5
x_0	2.6364	-0.1456	-0.0020	0.5695	0.3986	0.1628
x_1	-0.1456	2.3292	0.1064	0.2074	0.7591	0.1475
x_2	-0.0020	0.1064	2.4205	0.3489	0.1488	0.6143
x_3	0.5695	0.2074	0.3489	2.2351	-0.4355	-0.3277
x_4	0.3986	0.7591	0.1488	-0.4355	2.1625	-0.3445
x_5	0.1628	0.1475	0.6143	-0.3277	-0.3445	2.2518

# learn.save("17_jan_all_gaps")

Path('models/17_jan_all_gaps.pth')

train_show_save(learn, 2, 1e-3)

train_show_save(learn, 2, 1e-3)

epoch	train_loss	valid_loss	rmse	rmse_gap	r2	r2_gap	time
0	-149.288297	-152.804403	0.072760	0.179190	0.956827	-61340364986143763995601928192.000000	02:35
1	-160.099637	-165.002097	0.071263	0.181697	0.963587	-48387608947773502947627892736.000000	02:37

# learn.save("17_jan_all_gaps_final")

Path('models/17_jan_all_gaps_final.pth')

Results

learn.load("17_jan_all_gaps_final");

p = show_results(learn, control=hai_control, items = [items[i] for i in [0,5,4]], show_metric=False, units=list(units.values()))
p

p0 = show_results(learn, control=hai_control, items = [items[i] for i in [0]], n_cols=2, show_metric=False, props={'width': 350, 'height': 250},  units=list(units.values()))
p0

p1 = show_results(learn, control=hai_control, items = [items[i] for i in [5]], n_cols=2, show_metric=False, props={'width': 350, 'height': 250}, units=list(units.values()))
p1

p2 = show_results(learn, control=hai_control, items = [items[i] for i in [4]], n_cols=2, show_metric=False, props={'width': 350, 'height': 250}, units=list(units.values()))
p2

save_plot(p0, "results_gap_all_vars_g0")
save_plot(p1, "results_gap_all_vars_g1")
save_plot(p2, "results_gap_all_vars_g2")

Path('/home/simone/Documents/uni/Thesis/GPFA_imputation/analysis/presentations/plots_18_jan/results_gap_all_vars_g2.vl.json')

interact_results(learn, hai, hai_era)

/home/simone/anaconda3/envs/data-science/lib/python3.10/site-packages/fastai/callback/core.py:69: UserWarning: You are shadowing an attribute (__class__) that exists in the learner. Use `self.learn.__class__` to avoid this
  warn(f"You are shadowing an attribute ({name}) that exists in the learner. Use `self.learn.{name}` to avoid this")

<function meteo_imp.kalman.fastai.interact_results.<locals>._inner(gap_len, items_idx, control_lags, block_len, shift, **var_names)>

Gap in only 1 variable

dls2 = imp_dataloader(hai, hai_era, var_sel = ['TA'], block_len=200, gap_len=10, bs=20, control_lags=[1], n_rep=2)

model2 = init_smart(3,3, hai, pca=True).cuda()

model2

state	x_0	x_1	x_2	x_3	x_4	x_5
x_0	1.0000	0.0000	0.0000	1.0000	0.0000	0.0000
x_1	0.0000	1.0000	0.0000	0.0000	1.0000	0.0000
x_2	0.0000	0.0000	1.0000	0.0000	0.0000	1.0000
x_3	0.0000	0.0000	0.0000	1.0000	0.0000	0.0000
x_4	0.0000	0.0000	0.0000	0.0000	1.0000	0.0000
x_5	0.0000	0.0000	0.0000	0.0000	0.0000	1.0000

trans cov (Q)

state	x_0	x_1	x_2	x_3	x_4	x_5
x_0	0.1000	0.0000	0.0000	0.0000	0.0000	0.0000
x_1	0.0000	0.1000	0.0000	0.0000	0.0000	0.0000
x_2	0.0000	0.0000	0.1000	0.0000	0.0000	0.0000
x_3	0.0000	0.0000	0.0000	0.1000	0.0000	0.0000
x_4	0.0000	0.0000	0.0000	0.0000	0.1000	0.0000
x_5	0.0000	0.0000	0.0000	0.0000	0.0000	0.1000

trans off

state	offset
x_0	0.0000
x_1	0.0000
x_2	0.0000
x_3	0.0000
x_4	0.0000
x_5	0.0000

obs matrix (H)

variable	x_0	x_1	x_2	x_3	x_4	x_5
y_0	0.0168	0.9301	-0.3669	0.0000	0.0000	0.0000
y_1	0.9998	-0.0198	-0.0045	0.0000	0.0000	0.0000
y_2	0.0114	0.3667	0.9303	0.0000	0.0000	0.0000

obs cov (R)

variable	y_0	y_1	y_2
y_0	0.0100	0.0000	0.0000
y_1	0.0000	0.0100	0.0000
y_2	0.0000	0.0000	0.0100

obs off

variable	offset
y_0	0.0000
y_1	0.0000
y_2	0.0000

contr matrix (B)

state	c_0	c_1	c_2	c_3	c_4	c_5
x_0	-0.0168	-0.9998	-0.0114	0.0168	0.9998	0.0114
x_1	-0.9301	0.0198	-0.3667	0.9301	-0.0198	0.3667
x_2	0.3669	0.0045	-0.9303	-0.3669	-0.0045	0.9303
x_3	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
x_4	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
x_5	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000

init state mean

state	mean
x_0	0.0000
x_1	0.0000
x_2	0.0000
x_3	0.0000
x_4	0.0000
x_5	0.0000

init state cov

state	x_0	x_1	x_2	x_3	x_4	x_5
x_0	3.0000	0.0000	0.0000	0.0000	0.0000	0.0000
x_1	0.0000	3.0000	0.0000	0.0000	0.0000	0.0000
x_2	0.0000	0.0000	3.0000	0.0000	0.0000	0.0000
x_3	0.0000	0.0000	0.0000	3.0000	0.0000	0.0000
x_4	0.0000	0.0000	0.0000	0.0000	3.0000	0.0000
x_5	0.0000	0.0000	0.0000	0.0000	0.0000	3.0000

loss2 = loss_func=KalmanLoss(only_gap=False)
learn2 = Learner(dls2, model2, loss2, cbs=[Float64Callback], metrics=imp_metrics)

train_show_save(learn2, 1, 1e-3)

epoch	train_loss	valid_loss	rmse	rmse_gap	r2	r2_gap	time
0	-60.562467	-78.105121	0.041354	0.095727	0.974146	-10.220484	03:00

train_show_save(learn2, 3, 1e-3)

epoch	train_loss	valid_loss	rmse	rmse_gap	r2	r2_gap	time
0	-99.062288	-107.225395	0.043633	0.077547	0.955371	-5.651017	02:56
1	-115.215551	-122.628456	0.048706	0.075058	0.960905	-4.600422	02:52
2	-130.109601	-137.303035	0.054807	0.069069	0.966896	-3.014253	02:58

learn2.loss_func = KalmanLoss(only_gap=True)

train_show_save(learn2, 1, .5e-3)

epoch	train_loss	valid_loss	rmse	rmse_gap	r2	r2_gap	time
0	-2.358371	-2.383616	0.058387	0.077012	0.958465	-4.192017	01:52

train_show_save(learn2, 1, 2e-4)

epoch	train_loss	valid_loss	rmse	rmse_gap	r2	r2_gap	time
0	-2.387042	-2.445100	0.057919	0.073482	0.956432	-3.904405	01:51